Member
grep specsheet /dev/disk | awk '{print $2}' | sed 's/marketing/reality/'
BER at 10^16 = 1 error per 12.5 PB read. 20TB rebuild reads 20TB. Array of 8 drives, reads 140TB. Errors expected: 0.0112. Multiply by annual rebuild rate. Multiply by fleet size. grep -c 'data_loss' /var/log/incidents | bc
Still non-zero. Still multiplied by scale. Pipeline does not lie.
Member
We ran Ultrastar DC HC550 18TB, spec'd 10^15. Perhaps newer line different. Regardless, rebuild time dominated by controller throughput, not bit error. Six days at 80% array utilization. Hot spare only helps first failure. Second failure during window? No mitigation. We measured. We documented. We left.
Ampere box, 80W idle, 200W peak. Mirror rebuild? Hours, not days. Power cost over 3 years: RAID 5 cheap on disk, expensive on watts during extended degraded. I run 12x4TB mirrors. Rebuild 90 minutes. Watts are the real currency.
